NovaFabric — prove what a run actually did.
Tracing tells you what happened. NovaFabric tells you whether a past run can be replayed, compared, and proven — as a portable evidence capsule you own, produced in your own infrastructure with no accounts, no telemetry, and no changes to the code being observed.
What it does
NovaFabric turns any command — a script, an agent, a model call, an HPC training job, a notebook cell — into an execution capsule: a schema-valid, secret-redacted, replayable evidence folder written to your own disk. Four verbs do the work:
- Capture —
nova capture python my_agent.pyrecords the run as a capsule. The observed program is not modified; there is no SDK to import and no decorator to add. - Validate — a capsule is checked against a published schema, so “I have the evidence” is a testable statement rather than a filing habit.
- Replay — re-run a capsule in one of four modes (forensic, mocked, semantic, exact), including offline and without API keys.
- Diff & trace lineage — compare two runs field by field, and follow assets, runs and outputs across a chain of work.
- Version0.101.0
- Released7 August 2026
- LicenceApache-2.0
- StatusBeta
- RuntimePython ≥ 3.12
- DeploymentSelf-hosted · laptop to cluster
- SourceGitHub
- PackagePyPI
- Docsnovafabric.ai
Accountability for an AI system means someone can check it afterwards — without your permission.
Most of what is currently sold as AI governance is a dashboard: a live view, hosted by a vendor, of rows in that vendor's database. It answers “is the system up?” well. It answers “can you prove, in two years, what this model was given and what it returned?” badly — because the record lives somewhere you do not control, in a format only that vendor can read, for as long as that vendor chooses to exist.
NovaFabric inverts the unit of value. The artifact is a folder on your disk, in a documented schema, with a manifest you can hash. An evidence bundle verifies with sha256sum and an ed25519 verifier — no NovaFabric installation required to check it. That is the structural version of the claim: the ability to audit the run does not depend on the auditor trusting, licensing, or even having heard of the tool that produced it.
Built to interoperate rather than to capture
Capsules speak existing standards — OpenTelemetry GenAI semantic conventions, in-toto/DSSE, RFC 3161 timestamping, W3C PROV, OpenLineage and CycloneDX — rather than a private format. There is no account to create, no telemetry call home, and nothing that stops working when a network is absent, which is also what makes it usable on an air-gapped cluster.
What NovaFabric deliberately is not.
Being specific about what a system cannot do is part of what makes claims about what it can do worth anything. NovaFabric is not the right tool if you need:
- Real-time monitoring and alerting. This is stated in the project's own README. NovaFabric produces evidence after the fact; if what you want is a live dashboard that pages you at 3am, use an observability platform and keep NovaFabric for the questions that arrive months later.
- A hosted service. There is no SaaS. It runs in your infrastructure, which means you own the storage, the retention policy and the operational burden along with the evidence.
- A guarantee that a replay reproduces the original result. A capsule records what a run did and lets you re-execute it; genuinely non-deterministic systems stay non-deterministic. The four replay modes exist precisely because “reproduce it exactly” is not always achievable, and the mode you used is recorded rather than glossed over.
- A compliance certification. Evidence capsules can support an audit. They are not a regulatory approval, and no tool can supply one.
Version 0.101.0 is published as beta. It is versioned on a public registry and its source is readable, so the maturity claim is one you can check rather than one you have to accept — but it is beta, and this page will not describe it otherwise.
Use it, or help build it.
NovaFabric is free and open source under Apache-2.0, developed in the open. Issues, integrations, replay-mode edge cases and interoperability testing against the standards above are all welcome.
Open the repository